Pew Research examined nearly half a million English‑language pages harvested from the Common Crawl archive, covering a period that starts roughly two years prior to ChatGPT’s November 2022 release and extends to mid‑2026. To estimate AI involvement, they applied Open Pangram’s classifier, which flags pages when the model’s confidence exceeds a preset threshold for synthetic text. The approach treats any page above that threshold as likely authored or substantially edited by an AI system.
In a July 2026 random draw of 10,000 pages, approximately 10 % surpassed the classifier’s confidence cutoff, signalling strong AI‑authorship traits. Because the draw inevitably contained pages created before generative models existed, Pew removed all entries dated earlier than ChatGPT’s launch. Re‑evaluating the filtered set revealed that over one‑third (35 %) of the post‑ChatGPT pages exhibited detectable AI markers, suggesting that roughly a third of newer web content bears measurable AI influence.
- Sample: ~500k pages from Common Crawl
- Detection: Open Pangram AI‑text classifier
- Timeframe: pre‑ChatGPT to July 2026
- Post‑ChatGPT AI‑authored share: 35 %
Why this matters
Source facts show that 35 % of web pages published after ChatGPT’s release carry detectable AI‑text signatures, based on Open Pangram’s classifier applied to a Common Crawl sample. This indicates a growing reliance on generative models for content creation, which may dilute human editorial oversight and increase the likelihood of recurrent AI‑to‑AI consumption cycles. Inferentially, if the trend continues, the proportion of synthetic text could surpass half of new web output within a few years, raising concerns about source attribution, misinformation risk, and the need for robust provenance tools.
Share this article
Found this insightful? Share it with your community on Reddit, X, or copy the link.